When hair breaks a lesion border, repair the high frequencies before decoding: 89.69% Dice on ISIC 2018
Synopsis
In dermoscopic images, hair occlusion, uneven illumination and weak local contrast fragment lesion boundaries, while self-attention tends to emphasize lower-frequency content and leaves those weak borders underrepresented in skip connections. Swin-WaveNet decomposes each encoder skip feature with a wavelet transform, applies selective state-space scanning only to the high-frequency subbands to reconnect interrupted contours, gates shallow high-frequency responses with deeper semantic guidance to suppress hair and artifact structures, and sharpens boundaries during upsampling with high-frequency attention. On ISIC 2018 it reached a mean Dice of 89.69% and HD95 of 23.60 pixels over three runs, close to Swin-Unet, and 91.69% Dice on ISIC 2017 without fine-tuning.
Interpretation
Skip features for skin lesion segmentation are no longer passed straight to the decoder; before decoding they are split by frequency, repaired and filtered so that broken contours are reconnected before reconstruction. Earlier segmentation networks learned boundary cues in the spatial domain, where useful edge evidence is entangled with gradients from hair, reflections and noise, and skip connections carried both forward. On ISIC 2018 the complete model reached a three-run mean Dice of 89.69%±0.51% and HD95 of 23.60±1.04 pixels, the numerically highest Dice and lowest HD95 among the compared methods but only slightly different from Swin-Unet; in a single-run component ablation the complete model raised Dice by 0.96 percentage points and lowered HD95 by 1.16 pixels relative to the internal baseline.
Each skip branch is split by a parameter-free Haar wavelet transform into one low-frequency approximation band and three directional detail bands; the low-frequency band is held constant and only the concatenated high-frequency stream is modeled. A two-dimensional selective scan unfolds the feature map into four one-dimensional sequences along different directions, letting information propagate recursively across short interruptions and reconnect weakened edge responses; a learnable scale controls the magnitude of the high-frequency correction, which is recombined with the approximation band by an inverse wavelet transform. In the scanning-target comparison, scanning only the high-frequency subbands gave Dice 89.70% and HD95 23.63 pixels, better than 24.50 pixels for scanning the full feature map and 26.45 pixels for scanning only the low-frequency branch, though the high-frequency variant adds 2.61 million parameters and 0.28 GFLOPs from concatenating three subbands.
Perspective
Evaluation covers three public datasets: ISIC 2018 as the primary benchmark with 2594 training, 100 validation and 1000 test images; direct testing on the 600-image ISIC 2017 test set after training on ISIC 2018 without fine-tuning; and the 1000 endoscopic images of Kvasir-SEG on a fixed 7:1:2 split. All models were compared in one codebase under a unified protocol with 384×384 inputs, a fixed 0.5 threshold and no postprocessing, reporting mean and standard deviation over three runs with seeds 3232, 3407 and 42. Intended for readers interested in skip-connection design and boundary handling in medical image segmentation.
Diffuse lesions with gradual pigmentation changes remain difficult because their lesion-to-background transitions contain limited high-frequency evidence, leaving the high-frequency selective state-space module and the high-frequency attention decoder with less structural information. Component and scanning-target ablations are single-run, and the authors note the high-frequency variant has more parameters, so the scanning target cannot be fully isolated; the cross-dataset behavior is attributed speculatively to frequency decomposition partly separating appearance variation from structural detail. Evaluation is limited to three public datasets and three training runs, and the authors call for multicenter validation on data from different imaging devices and for a lighter encoder for resource-constrained deployment.